Complex crystal structure prediction using ML-enhanced multi-minima iterative genetic algorithm

Cai-Zhuang Wang; Ling Tang; Paul C. Canfield; Tyler J. Slade; Weiyi Xia

arxiv: 2607.01004 · v1 · pith:727J5ZNBnew · submitted 2026-07-01 · ❄️ cond-mat.mtrl-sci

Complex crystal structure prediction using ML-enhanced multi-minima iterative genetic algorithm

Ling Tang , Weiyi Xia , Tyler J. Slade , Paul C. Canfield , Cai-Zhuang Wang This is my paper

Pith reviewed 2026-07-02 09:28 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci

keywords crystal structure predictiongenetic algorithmmachine learning interatomic potentialternary compoundsLa-Co-Pb systemmetastable phasesmaterials discovery

0 comments

The pith

A multi-minima iterative genetic algorithm combined with a machine learning interatomic potential predicts complex crystal structures in ternary systems using only composition data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a multi-minima iterative genetic algorithm that pairs an artificial neural network machine learning potential with an iterative penalty scheme inspired by metadynamics. This combination explores intricate energy landscapes in systems where elements do not mix easily, such as the La-Co-Pb ternary. The method identifies the novel Pbam ground-state structure of La4Co4Pb and exactly reproduces the orthorhombic structure of La5CoPb2 as confirmed by x-ray diffraction. It succeeds without drawing on existing structural databases and locates both the lowest-energy phase and competing metastable states. The result demonstrates a route to finding entirely new structure types for quantum and magnetic materials.

Core claim

The MMIGA approach integrates an ANN-ML interatomic potential with a metadynamics-inspired penalty scheme inside a genetic algorithm framework. When applied to the La-Co-Pb system, it locates the Pbam structure of La4Co4Pb and the orthorhombic structure of La5CoPb2, matching independent x-ray diffraction data using only the input composition.

What carries the argument

The multi-minima iterative genetic algorithm (MMIGA) that employs an artificial neural network machine learning interatomic potential together with an iterative penalty scheme to sample multiple minima on the energy surface.

If this is right

The method locates both the global minimum and relevant metastable phases within the same run.
It enables discovery of structure types absent from existing databases.
The approach supplies theoretical guidance on which phases are likely to form in antagonistic-pair systems.
It operates with only the chemical composition as input for ternary compounds.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same penalty-enhanced genetic search could be tested on quaternary compositions with comparable immiscibility.
Retraining the neural-network potential on a broader set of calculated energies might extend reliable predictions to other rare-earth transition-metal systems.
Mapping metastable states this way could inform experimental synthesis routes that stabilize otherwise overlooked phases.

Load-bearing premise

The artificial neural network interatomic potential, trained on available data, correctly captures the energy landscape of the La-Co-Pb system even though cobalt and lead do not mix.

What would settle it

Prediction of a structure for a new La-Co-Pb composition or similar ternary that differs from the structure later measured by x-ray diffraction on the synthesized compound.

read the original abstract

Current machine learning (ML) approaches for materials discovery rely heavily on known structural databases, limiting their ability to identify entirely novel structure types. In this work, we develop a multi-minima iterative genetic algorithm (MMIGA) that integrates an artificial-neural-network machine learning (ANN-ML) interatomic potential with an iterative, metadynamics-inspired penalty scheme. We demonstrate the robustness of this method on a complex ternary La-Co-Pb system, characterized by Co-Pb immiscibility and an intricate energy landscape. The ML-enhanced MMIGA successfully predicts the ground-state Pbam structure of the recently synthesized La4Co4Pb antagonistic-pair-phase, a novel structure missed by previous database-reliant ML predictions, while also identifying multiple metastable competing phases. Additionally, we challenged the MMIGA method to predict the structure of La5CoPb2 antagonistic-pair-phase, a new compound discovered during earlier attempts to synthesize the predicted phase La3CoPb. With only knowledge of the composition, our MMIGA approach successfully predicts the orthorhombic structure of La5CoPb2, producing an exact match with the structure independently determined by x-ray diffraction. By efficiently mapping both global minimum and relevant competing metastable states, this approach provides critical theoretical insights into phase selection for novel quantum and magnetic materials.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

MMIGA with the ANN-ML potential recovers the exact experimental structure of La5CoPb2 from composition alone, which is the concrete result worth noting, though the potential's fidelity for this immiscible system remains the open question.

read the letter

The headline result is that the method found the orthorhombic Pbam structure of La5CoPb2 exactly as later determined by x-ray diffraction, using only the composition as input. It also recovered the La4Co4Pb structure that database-based ML approaches had missed. That match gives the claim a clear anchor.

The integration of the ANN-ML potential with an iterative penalty inside the genetic algorithm is the part presented as new. It lets the search map both the ground state and several metastable phases in one go, which is useful for systems like this ternary where Co-Pb immiscibility creates a complicated landscape.

The soft spot is the potential. The abstract itself highlights the intricate energy surface and immiscibility, yet the supplied text gives no information on how many ternary configurations were used in training, what validation was done against DFT, or how the energy ordering of the reported phases was checked. Without those details it is hard to tell whether the success reflects accurate physics or a fortunate fit for these particular cases.

This paper is aimed at people working on evolutionary structure prediction who need to move beyond database-trained models. A reader who already uses genetic algorithms or ML potentials will see a concrete example of how the penalty scheme helps explore multiple minima.

It deserves peer review because the experimental match is independent and the method is laid out enough to test. The potential details can be requested in revision.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces a multi-minima iterative genetic algorithm (MMIGA) that couples an ANN-ML interatomic potential with an iterative, metadynamics-inspired penalty scheme. It claims that, given only composition, the method recovers the experimental Pbam structure of La4Co4Pb (a novel structure missed by database-reliant ML) and produces an exact match to the independently determined orthorhombic structure of La5CoPb2 in the La-Co-Pb system, while also locating multiple metastable phases.

Significance. If the ANN-ML potential is shown to be reliable, the approach would be significant for discovering novel structures in systems with immiscibility and intricate landscapes without reliance on existing databases. The validation against independent x-ray structures (rather than quantities derived from the same fitted potential) is a strength that reduces circularity. The focus on both global minimum and competing metastable states is also valuable for phase-selection insights.

major comments (2)

[Abstract] Abstract: the claim of an 'exact match' to the x-ray structure of La5CoPb2 (and the successful prediction of La4Co4Pb) is presented without any information on the training data used for the ANN-ML potential, whether ternary La-Co-Pb configurations were included, or any cross-validation of energy ordering against DFT or experiment. This information is load-bearing for the central claim that the surrogate potential correctly ranks the true ground state amid Co-Pb immiscibility.
[Abstract] Abstract: no details are supplied on the number of independent MMIGA runs performed, the convergence criteria for the iterative penalty scheme, or how the method ensures that the reported structure is the global minimum rather than a local one. These omissions directly affect the robustness assertion for the La5CoPb2 prediction.

minor comments (1)

The abstract would benefit from a concise statement of the computational cost or number of force evaluations required per prediction to allow readers to assess practicality.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful review and constructive suggestions. We address each major comment below and have revised the manuscript to improve clarity on the points raised.

read point-by-point responses

Referee: [Abstract] Abstract: the claim of an 'exact match' to the x-ray structure of La5CoPb2 (and the successful prediction of La4Co4Pb) is presented without any information on the training data used for the ANN-ML potential, whether ternary La-Co-Pb configurations were included, or any cross-validation of energy ordering against DFT or experiment. This information is load-bearing for the central claim that the surrogate potential correctly ranks the true ground state amid Co-Pb immiscibility.

Authors: We agree that the abstract should be self-contained regarding the ANN-ML potential. The full manuscript (Methods and Results sections) details that the potential was trained on a DFT-generated dataset that explicitly includes ternary La-Co-Pb configurations sampled across relevant compositions, with energy ordering cross-validated against independent DFT calculations on held-out structures. To address the referee's concern directly, we have revised the abstract to include a concise statement summarizing the training data composition and validation against DFT. revision: yes
Referee: [Abstract] Abstract: no details are supplied on the number of independent MMIGA runs performed, the convergence criteria for the iterative penalty scheme, or how the method ensures that the reported structure is the global minimum rather than a local one. These omissions directly affect the robustness assertion for the La5CoPb2 prediction.

Authors: The full manuscript describes the MMIGA procedure, including multiple independent runs and the metadynamics-inspired penalty scheme for exploring distinct minima. We acknowledge that these parameters are not summarized in the abstract. We have added a brief clause to the abstract noting the use of repeated independent runs with the iterative penalty to identify the lowest-energy structure consistently recovered across runs, thereby strengthening the robustness claim without altering the technical content. revision: yes

Circularity Check

0 steps flagged

No significant circularity; central predictions validated against independent XRD structures

full rationale

The paper's headline result is that MMIGA + ANN-ML potential recovers the experimental Pbam structure of La5CoPb2 (and Pbam for La4Co4Pb) from composition alone, with explicit statements that the match is to 'the structure independently determined by x-ray diffraction.' No load-bearing self-citation, self-definitional loop, or 'fitted input called prediction' appears in the abstract or described method. The genetic search output is compared to external experimental data rather than to quantities derived from the same fit, satisfying the criterion for a self-contained result against external benchmarks. The fidelity of the underlying potential is a separate correctness question, not a circularity issue.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the accuracy of a fitted ANN-ML potential and the assumption that the penalty scheme plus genetic search sufficiently samples the energy landscape of an immiscible ternary system.

free parameters (1)

ANN-ML interatomic potential parameters
The potential is trained on data and therefore contains fitted parameters whose values are not reported in the abstract.

axioms (1)

domain assumption The metadynamics-inspired penalty scheme prevents the search from becoming trapped in a single minimum and allows discovery of competing metastable phases.
Invoked in the description of the MMIGA method.

pith-pipeline@v0.9.1-grok · 5773 in / 1242 out tokens · 32389 ms · 2026-07-02T09:28:57.412784+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

8 extracted references

[1]

converged

Introduction Novel advanced materials discovery and development is critical for driving technological innovation and sustainability. The vast compositional and structural landscape of inorganic and organic materials remains largely unexplored, offering a compelling opportunity for computational and experimental discovery of novel multi-element compounds w...
[2]

cut-and-paste

Search for new ternary La-Co-Pb compounds We demonstrated the robust performance of the MMIGA method for global structure prediction of complex compounds using ternary La-Co-Pb compounds containing the immiscible pair of Co-Pb elements as a test case. It was proposed by Canfield [14] that ternary intermetallic compounds formed by adding a third element to...

2023
[3]

converged

J. P. Perdew, K. Burke, and M. Ernzerhof, Generalized Gradient Approximation Made Simple, Phys. Rev. Lett. 77, 3865(1996). [32] Atsushi Togo, Laurent Chaput, Terumasa Tadano, and Isao Tanaka, Implementation strategies in phonopy and phono3py, J. Phys. Condens. Matter 35, 353001(2023). [33] Atsushi Togo, First-principles Phonon Calculations with Phonopy an...

1996
[4]

Metastable structures of La4Co4Pb from MMIGA search In the 1st iteration of the MMIGA for La4Co4Pb (Z = 4 f.u.), our study found two metastable structures with Pmm2 and C2/m space group symmetries respectively. The Co atoms in these two structures also form zigzag buckling layers, but the topology of the atomistic connection is different from the Kagome l...
[5]

converged

MMIGA searches for the La5CoPb2 phases with Z = 2 formula unites Fig. S4. The iterative GA searches for La5CoPb2 crystalline phase with Z = 2 formula units (16 atoms/cell). The ML potential energy of the lowest-energy structure in the GA pool and the average energy of the pool as function of GA generations in the 8 successive iterative GA searches are sho...
[6]

converged

MMIGA searches for the La5CoPb2 phases with Z = 3 formula unites Fig. S7. The iterative GA searches for La5CoPb2 crystalline phase with Z = 3 formula units (24 atoms/cell). The ML potential energy of the lowest-energy structure in the GA pool and the average energy of the pool as function of GA generations in the 8 successive iterative GA searches are sho...
[7]

Crystallography information Table S1. Anisotropic thermal displacement parameters for La5CoPb2 Atom U11 U22 U33 U23 U13 U12 Pb1 0.01598(13) 0.01176(12) 0.01605(12) -0.00056(5) 0.00068(6) -0.00017(5) La1 0.0199(2) 0.01367(19) 0.0176(2) 0 0.00389(15) 0 La2 0.0204(2) 0.01315(19) 0.01642(19) 0 0.00387(15) 0 La3 0.0160(2) 0.0137(2) 0.0261(2) 0 0.00195(15) 0 La...

1904
[8]

Physical property measurements Fig. S10. (a) The temperature dependent resistance and (b) magnetization (M/H) of La5CoPb2 measured along the b-axis (the long axis of the rodlike crystals). Both datasets indicate that the Co atoms are non-moment bearing in La5CoPb2, showing no signatures of a phase transition and behavior characteristic of a Pauli-paramagn...

1993

[1] [1]

converged

Introduction Novel advanced materials discovery and development is critical for driving technological innovation and sustainability. The vast compositional and structural landscape of inorganic and organic materials remains largely unexplored, offering a compelling opportunity for computational and experimental discovery of novel multi-element compounds w...

[2] [2]

cut-and-paste

Search for new ternary La-Co-Pb compounds We demonstrated the robust performance of the MMIGA method for global structure prediction of complex compounds using ternary La-Co-Pb compounds containing the immiscible pair of Co-Pb elements as a test case. It was proposed by Canfield [14] that ternary intermetallic compounds formed by adding a third element to...

2023

[3] [3]

converged

J. P. Perdew, K. Burke, and M. Ernzerhof, Generalized Gradient Approximation Made Simple, Phys. Rev. Lett. 77, 3865(1996). [32] Atsushi Togo, Laurent Chaput, Terumasa Tadano, and Isao Tanaka, Implementation strategies in phonopy and phono3py, J. Phys. Condens. Matter 35, 353001(2023). [33] Atsushi Togo, First-principles Phonon Calculations with Phonopy an...

1996

[4] [4]

Metastable structures of La4Co4Pb from MMIGA search In the 1st iteration of the MMIGA for La4Co4Pb (Z = 4 f.u.), our study found two metastable structures with Pmm2 and C2/m space group symmetries respectively. The Co atoms in these two structures also form zigzag buckling layers, but the topology of the atomistic connection is different from the Kagome l...

[5] [5]

converged

MMIGA searches for the La5CoPb2 phases with Z = 2 formula unites Fig. S4. The iterative GA searches for La5CoPb2 crystalline phase with Z = 2 formula units (16 atoms/cell). The ML potential energy of the lowest-energy structure in the GA pool and the average energy of the pool as function of GA generations in the 8 successive iterative GA searches are sho...

[6] [6]

converged

MMIGA searches for the La5CoPb2 phases with Z = 3 formula unites Fig. S7. The iterative GA searches for La5CoPb2 crystalline phase with Z = 3 formula units (24 atoms/cell). The ML potential energy of the lowest-energy structure in the GA pool and the average energy of the pool as function of GA generations in the 8 successive iterative GA searches are sho...

[7] [7]

Crystallography information Table S1. Anisotropic thermal displacement parameters for La5CoPb2 Atom U11 U22 U33 U23 U13 U12 Pb1 0.01598(13) 0.01176(12) 0.01605(12) -0.00056(5) 0.00068(6) -0.00017(5) La1 0.0199(2) 0.01367(19) 0.0176(2) 0 0.00389(15) 0 La2 0.0204(2) 0.01315(19) 0.01642(19) 0 0.00387(15) 0 La3 0.0160(2) 0.0137(2) 0.0261(2) 0 0.00195(15) 0 La...

1904

[8] [8]

Physical property measurements Fig. S10. (a) The temperature dependent resistance and (b) magnetization (M/H) of La5CoPb2 measured along the b-axis (the long axis of the rodlike crystals). Both datasets indicate that the Co atoms are non-moment bearing in La5CoPb2, showing no signatures of a phase transition and behavior characteristic of a Pauli-paramagn...

1993